Seedance 2.5 多角色对白回执审计

Use this skill after an authorized operator has generated a result and returned a redacted receipt plus manual observations. It separates technical receipt completeness from the human decision to release.

Prerequisites

• - Python 3.9 or newer.
• The companion contract JSON or an equivalent redacted cast contract.
• A JSON receipt containing task status and output metadata, plus one observation per character and a manual review object. Do not put media bytes or secret URLs in the receipt.

Inputs and outputs

Run python scripts/audit_result.py examples/pass_receipt.json --contract examples/pass_contract.json --out audit.json --report audit.md.

The auditor checks completed status, a non-empty output identifier, positive output bytes, a 64-character SHA256, observation coverage, and explicit human review fields for appearance, wardrobe, voice, timing, and content. It outputs a redacted audit with release_ready: false so an operator must make the final release decision.

Execution and acceptance

1. 1. Export the generation receipt without tokens, private URLs, personal data, or media.
2. Record honest observations after watching and listening in the buyer\'s own environment.
3. Run the command. Exit code 0 means the technical audit is complete; exit code 1 means it is blocked. Neither result proves quality or model consistency.

The audit reports identity drift, voice drift, early or late dialogue, wardrobe mismatch, and missing manual review as declared observations. It never downloads a file, retries a task, calls a model, or publishes a result.

Permissions and failure boundaries

Manual review must be performed by an authorized person with access to the actual output. Keep private media out of examples. If the receipt is incomplete or a reviewer flags drift, hold the output and regenerate only after an explicit human choice and a fresh cost check.

Standalone use

This skill can be used independently to gate a completed receipt, even when the cast-contract skill was created elsewhere. It is the final evidence step in this two-skill workflow, not an automatic approval.